CN103234916B - Prediction method for net photosynthetic rate of population - Google Patents

Abstract

The invention discloses a group net photosynthetic rate prediction method, aiming at overcoming the difficulties in establishing a group net photosynthetic rate prediction regression equation, establishing a group net photosynthetic rate prediction model that requires more relevant data and low prediction accuracy. The steps are as follows: 1. Obtain the spectral radiation ratio data of each band of visible light in the region: install a portable multispectral radiometer at a height H higher than the plant canopy in the test area with an area of S; obtain the visible light in each band of the test area at different times Spectral radiation ratio data M_D is normalized to [0, 1]. 2. Obtain the net photosynthetic rate data of the population; 3. Construct the bionic kernel function; 4. Establish the SVM training set and prediction set; 5. Establish the tool selection and parameter optimization of the prediction model; 6. Predict the net photosynthetic rate of the population: obtain the prediction model model; get the predicted value predict of the net photosynthetic rate of the population and judge the reliability of the prediction model model.

Description

A Prediction Method of Population Net Photosynthetic Rate

technical field

The invention relates to a method for predicting the net photosynthetic rate of plant populations, more specifically, the invention relates to a method for predicting the net photosynthetic rate of the population with a bionic kernel function.

Background technique

As the most necessary resource for plants, light is an important factor affecting its form and function. The individual net photosynthetic rate reflects the accumulation of organic matter in a single plant and is an important factor affecting the form and function of a single plant. The group net photosynthetic rate is The sum of the individual net photosynthetic rates of this type of plant in a region reflects the accumulation of total photosynthetic organic matter of the plants in the region over a period of time, which is of great reference value for analyzing the overall shape and function of the plants in the region. Speed has a strong practical significance in agricultural production. Since the size of the leaf area reflects the size of the net photosynthetic rate, we can predict the net photosynthetic rate by establishing a regression equation between the leaf area and the net photosynthetic rate, but the traditional method of testing the leaf area has limitations such as complexity and expensive equipment sex. If the regression equation method between leaf area and net photosynthetic rate is used to ensure the accuracy of the predicted value of net photosynthetic rate, a large amount of sample data is required. To predict the net photosynthetic rate of the population, it is necessary to measure the leaf area of a large number of plants, and the workload more onerous.

Support Vector Machine (Support Vector Machine, SVM) is a machine learning method based on statistical learning theory. It shows many unique advantages in solving small samples, nonlinear and high-dimensional pattern recognition. In SVM, the kernel function is Its core part transforms the inner product operation of high-dimensional space into the calculation of kernel function of low-dimensional input space, which solves the problem of "dimension disaster" calculated in high-dimensional feature space, and the change of its form and parameters will implicitly It can change the mapping from input space to feature space, and then affect the properties of feature space. The kernel function is essentially an inner product. Its basic function is to accept vectors in two low-dimensional spaces, and to calculate the vector inner product value in a high-dimensional space after a certain transformation, that is, to determine the low-dimensional space to the high-dimensional space. The mapping relationship becomes the key to solving linear inseparability.

Through the study of the colorful chameleon known as the incomparably gorgeous rainbow in the biological world, it is found that its skin color not only changes with the color of the overall environment, but also the local skin color is similar to the color of its close environment, and can also change with the color of the local environment. It changes accordingly, showing a strong ability to adapt to the global surrounding environment color and local surrounding color. Starting from the natural phenomenon that chameleons adapt to the surrounding environment, it is of great significance to construct a bionic kernel function that can be adjusted independently for global data and local data.

Contents of the invention

The technical problem to be solved by the present invention is to overcome the relatively large workload of plant leaf area in the test area existing in the prior art, the difficulty of establishing a population net photosynthetic rate prediction regression equation, and the establishment of a population net photosynthetic rate prediction model. The problem of low accuracy provides a method for predicting the net photosynthetic rate of populations.

In order to solve the above-mentioned technical problems, the present invention is realized by adopting the following technical scheme: the steps of the described method for predicting the net photosynthetic rate of a group are as follows:

1) Obtain the spectral radiation ratio data of each band of visible light in the region:

(1) Install a portable multispectral radiometer model MSR-16 at a height H higher than the plant canopy in the test area with an area S, wherein: S takes a value of 6m ² to 55m ² , and H takes a value of 1.2m or 2m, the effective measurement diameter of the portable multispectral radiometer model MSR-16 on the ground directly below it is H/2, and the observation area is S ₁ , S ₁ =[π×(H/4) ² ], calculated The number of test units divided by the test area is M, and its value is the integer part of S/S ₁ ;

(2) In the specified measurement time period, measure once every hour, randomly select N test units each time, where N<M, and require that the test units selected each time are different from the test units selected at other measurement times, each time When measuring, place the portable multi-spectral radiometer of MSR-16 at H above the test unit during the measurement. The average value of 5 fixed points is used as the spectral composition of the test unit, and the average value of 3 test units is used as the data of the spectral radiation ratio relationship of each band of visible light in the test area during this period;

(3) According to the method of step (2), within the time span of the test, obtain the spectral radiation ratio data M_D of each band of visible light in the test area in different periods, and perform [0, 1] normalization processing to obtain normalization The transformed data M_D ₁ .

2) Obtain the population net photosynthetic rate data:

(1) The model of CI-310 portable photosynthesis measurement system produced by U.S. CID Company was adopted to measure the individual net photosynthetic rate;

(2) Randomly select 3 plants of the same variety in each selected test unit. If the number of plants is less than 3, measure according to the actual number of plants. Each plant randomly selects 5 leaves. If the number of leaves is less than 5 , measured according to the actual number of leaves, and the net photosynthetic rate of each leaf is measured once with a portable photosynthesis measurement system model CI-310, and the average value is taken as the net photosynthetic rate of the plant, and the average value of the three test units is used as the test period. The net photosynthetic rate of the plant population in the area;

(3) According to the method of step (2), within the test time span, obtain the net photosynthetic rate C_D of the plant population in the test area in different periods, carry out [0,1] normalization processing, and obtain the normalized Data C_D ₁ .

3) Build a bionic kernel function:

(1) Based on Gaussian kernelK ₁ (x, _xi )=exp(-γ||xx _i || ² ) and polynomial kernelK ₂ (x, _xi )=k(<x, _xi >+c) Kernel function, where k, γ and c are parameters, x and _xi are multi-dimensional vectors in low-dimensional space, input K ₁ (x, _xi )=exp(-γ||xx _i || ² ) into libsvm, As the SVM kernel function, train the training set data, use grid-search to search for the best parameters, and determine the γ value with the highest prediction accuracy for the training set, which is defined as the Γ value.

(2) Find out the maximum value diff(K ₁ (x _m , _xi )) and the minimum value diff(K ₁ (x _n , _xi )) of the tangent slope of the Gaussian kernel characteristic curve under the Γ value, and the corresponding tangent point Coordinates (x _m , K ₁ (x _m , _xi )), (x _n , K ₁ (x _n , _xi )), this type of tangent point is called a vertices.

(3) According to the corner coordinates (x _m ，K ₁ (x _m , _xi )), (x _n ,K ₁ (x _n , _xi )) and slope diff(K ₁ (x _m , _xi )) /ξ, diff(K ₁ (x _n , _xi ))/ξ, where ξ is a real number variable, determine the polynomial kernel expression, the expression is K ₂ '(x, _xi )=diff(K ₁ (x _m , _xi ))/ξ/ _xi ×(<x, _xi >+c) and K ₂ ″(x, _xi )＝diff(K ₁ (x _n ,xi ₎ )/ξ/xi _× (<x, _xi >+c) consists of two parts.

(4) Build a bionic kernel function:

K _bsf (x,x _i )＝exp(-Γ||xx _i || ² )+diff(exp(-Γ||x _B -x _i || ² )/ξ/xi _× (<x,x _i >+c)

Where: when ||x||>||x _i ||, x _B is taken as x _n , when ||x||<||x _i ||, x _B is taken as x _m , ||x||＝ When ||x _i ||, classify x and _xi as the same class, and do not need to use the kernel function to calculate it.

The bionic kernel function K _bsf (x, _xi ) can adjust the global data and local data through the adjustment of parameters Γ and ξ, so as to meet the requirements of different data classification.

4) Establish SVM training set and prediction set:

Divide data M_D ₁ into two parts of data M_D ₁₁ and M_D ₁₂ according to different time periods, divide data M_D ₁ into two parts of data M_D ₁₁ and M_D ₁₂ according to the time period, and divide data C_D ₁ into two parts of data C_D according to corresponding time periods ₁₁ and C_D ₁₂ , the first part of data M_D ₁₁ and C_D ₁₁ is used as the training set of SVM, and the second part of data M_D ₁₂ and C_D ₁₂ is used as the prediction set of SVM.

5) Tool selection and parameter optimization for building a predictive model.

6) Predict the net photosynthetic rate of the population.

The tool selection and parameter optimization for building a predictive model described in the technical proposal refers to:

Using MATLAB to put the bionic kernel function K _bsf (x, _xi ) into the libsvm toolbox to replace the position of the original RBF kernel function, realize the grid-search method to optimize parameters through the function SVMcgForRegress(), and use the grid-search method for preliminary optimization Parameter process, using Gaussian kernelK ₁ (x, _xi ) in the function SVMcgForRegress(), through [bestmse,bestc,bestg]=SVMcgForRegress(C_D ₁₁ ,M_D ₁₁ ,-8,8,-8,8), you can get Parameters bestc and bestg, bestc is the best penalty parameter c value, bestg is the best parameter γ value, that is, Γ value;

_The bionic kernel function K _bsf (x, _{xi )} is used in the function SVMcgForRegress(), and the best The best parameter bestCMG is the value of ξ.

The steps of predicting the population net photosynthetic rate described in the technical scheme are as follows:

1) Obtain the prediction model model

Using svmtrain() in the libsvm toolbox, that is, through model=svmtrain(C_D ₁₁ , M_D ₁₁ ,cmd), the prediction model model can be obtained, where cmd=['-c',num2str(bestc),'-g', num2str(bestg),'-s 3-p 0.01-t 2'], s is set to 3 to represent the e-SVR formula, p is set to 0.01 to set the value of the loss function in e-SVR, t is set to 2 to represent SVM The kernel function used in is K _bsf (x, _xi ); the model contains information about the type of kernel function, the number of support vectors and the coefficient range of support vectors in the decision function.

2) Obtain the predicted value predict of the net photosynthetic rate of the population

Using svmpredict() in the libsvm toolbox, through [predict,mse]=svmpredict(C_D ₁₂ ,M_D ₁₂ ,model) to realize the prediction of C_D ₁₂ using M_D ₁₂ , and obtain the predicted value predict of the net photosynthetic rate of the population.

3) Judging the reliability of the prediction model model

At the same time, the correlation coefficient R between the two is calculated through C_D ₁₂ and predict, and the accuracy of the predicted value predict is judged, that is, the reliability of the prediction model model.

Compared with prior art, the beneficial effects of the present invention are:

1. a kind of group net photosynthetic rate prediction method of the present invention compares with traditional method of setting up group net photosynthetic rate prediction regression equation, has model establishment process simple, required input data is few, prediction accuracy rate is high, easy to realize features.

2. A kind of population net photosynthetic rate prediction method of the present invention is compared with existing method for predicting individual net photosynthetic rate, forecasting population net photosynthetic rate can reflect the net photosynthetic rate situation of a region, is conducive to reflecting regional plant overall form and function.

3. Compared with the existing method of utilizing SVM to predict the net photosynthetic rate of a population, a kind of group net photosynthetic rate prediction method of the present invention adopts a novel kernel function, which realizes a better classification ability for SVM input data, and has The prediction model has the characteristics of strong generalization ability and high prediction accuracy.

4. A method for predicting the net photosynthetic rate of the population according to the present invention can realize the prediction of the net photosynthetic rate of the population only through the spectral radiation ratio relationship of each band of visible light in the region, and has the characteristics of less required input data, which is conducive to reflecting the visible light in the region The relationship between the ratio of spectral radiation in each band and the net photosynthetic rate.

5. A kind of group net photosynthetic rate prediction method of the present invention has adopted novel kernel function, has overcome the deficiency that the global learning ability of traditional kernel function and local learning ability can not be adjusted independently, can be according to the spectral radiation ratio relation of each band of visible light The characteristics of the data distribution adjust the performance of the kernel function, which improves the efficiency and accuracy of the forecasting process, and this method can be directly used in other forecasting applications.

6. A method for predicting net photosynthetic rate of a population according to the present invention is simple, convenient, fast, clear in steps, time-saving and labor-saving, and the prediction accuracy of the prediction method is high.

Description of drawings

Below in conjunction with accompanying drawing, the present invention will be further described:

Fig. 1 is the MSR-16 type portable multi-spectral radiometer placement position and the effective test area schematic diagram adopted in a kind of group net photosynthetic rate prediction method of the present invention;

Fig. 2 adopts the schematic diagram of CI-310 instrument test net photosynthetic rate for a kind of population net photosynthetic rate prediction method of the present invention;

Fig. 3 is the process block diagram that the grid-search method in a kind of group net photosynthetic rate prediction method of the present invention carries out preliminary optimization to parameter;

Fig. 4 is the process block diagram that the bionic kernel function parameter in a kind of group net photosynthetic rate prediction method of the present invention is determined;

Fig. 5 is a data distribution diagram of the spectral radiation ratio relationship of each band of visible light in the region after normalization by a population net photosynthetic rate prediction method according to the present invention;

Fig. 6 is the population net photosynthetic rate data distribution diagram after adopting a kind of population net photosynthetic rate prediction method normalized according to the present invention;

Fig. 7 is the training group collective net photosynthetic rate prediction curve (the first) of a kind of group net photosynthetic rate prediction method of the present invention;

Fig. 8 is the prediction group net photosynthetic rate prediction curve (the second) of a kind of group net photosynthetic rate prediction method of the present invention;

Fig. 9 is a flowchart of a method for predicting net photosynthetic rate of a population according to the present invention.

Detailed ways

The present invention is described in detail below in conjunction with accompanying drawing:

Referring to Fig. 9, the present invention overcomes the problems existing in the prior art, and provides a method for predicting the net photosynthetic rate of the population, that is, only using the spectral radiation ratio relationship of each band of visible light to realize the prediction of the net photosynthetic rate of the population with high accuracy method, the steps are as follows:

1. Obtain the ratio relationship data of spectral radiation in each band of visible light in the region

The steps to obtain the spectral radiance ratio data of each band of visible light in the region are as follows:

A portable multispectral radiometer of model MSR-16 produced in the United States was used to analyze the spectral composition and proportion changes of the visible light in the 460-710nm band incident above the portable multispectral radiometer.

1) Referring to Figure 1, assume that the area of the test area is S, and the value of S is 6m ² to 55m ² , the height of the portable multispectral radiometer model MSR-16 above the plant canopy is H, and the value of H is 1.2m or 2m, the effective measurement diameter of the portable multispectral radiometer model MSR-16 on the ground directly below it is H/2, and the observation area is S ₁ , S ₁ =[π×(H/4) ² ], where: π is the circumference ratio. According to the test area S divided during the test and the MSR-16 instrument test height H, the number of test units divided by the test area to be tested is calculated as M, and the value is taken as the integer part of S/S ₁ .

2) In the specified measurement time period, measure once every hour, randomly select N test units each time, where N<M, and require that the test units selected each time are different from the test units selected at other measurement times, each measurement , place the MSR-16 portable multispectral radiometer at H above the test unit during the measurement, and the value of H is 1.2m or 2m. Take 5 fixed-point measurements in each test unit, and measure each point twice. Taking the average, the average value of 5 fixed points is used as the spectral composition of the test unit, and the average value of 3 test units is used as the data of the spectral radiation ratio relationship of each band of visible light in the test area during this period.

3) According to the method of step 2), within the time span of the test, the data M_D of the spectral radiation ratio relationship of each band of visible light in the test area at different time periods are obtained, and the [0, 1] normalization process is performed to obtain the normalized The data M_D ₁ .

2. Obtain population net photosynthetic rate data

Referring to Fig. 2, the steps of obtaining the population net photosynthetic rate data method are as follows:

1) Use the model CI-310 portable photosynthesis measurement system produced by the American CID company to measure the individual net photosynthetic rate:

Select a 25×25 (cm ² ) square leaf chamber to insert into the host, press the probe at the front end of the sensor into the leaf chamber, insert the plug at the rear end of the sensor into the corresponding jack of the host, connect one end of the data cable to the host, and the other end of the data cable to the PC The PC card is connected to the computer through the PC card slot, and a tube is drawn from the "INTAKE" port of the CI-310 sensor to a 4L buffer bottle, which is placed in the outside air to obtain CO that is not affected by breathing. ₂ Air, _CO2 zeroing with the supplied soda lime tube before each measurement. Put the blade into the leaf chamber and close the leaf chamber, press the measurement switch of the main engine to start the measurement of the first leaf. A blade measurement.

2) Randomly select 3 plants of the same variety in each selected test unit (if the number of plants is less than 3, measure according to the actual number of plants), and randomly select 5 leaves for each plant (if the number of leaves is less than 5 , measured according to the actual number of leaves), the net photosynthetic rate of each leaf is measured once with a portable photosynthesis measurement system model CI-310, and the average value is taken as the net photosynthetic rate of the plant, and the average value of the three test units is used as the period. The net photosynthetic rate of the plant population in the test area.

3) According to the method of step 2), within the test time span, the net photosynthetic rate C_D of the plant population in the test area at different time periods is obtained, and the [0, 1] normalization process is performed to obtain the normalized data C_D ₁ .

3. Construct bionic kernel function

1) Take Gaussian kernelK ₁ (x, _xi )=exp(-γ||xx _i || ² ) and polynomial kernelK ₂ (x, _xi )=k(<x, _xi >+c) as the benchmark kernel function, where γ and c are parameters, and x and _xi are multidimensional vectors in low-dimensional space. Input K ₁ (x, _xi )=exp(-γ||xx _i || ² ) into libsvm as the SVM kernel function to train the training set data, use grid-search to search for the best parameters, and determine the pair The γ value with the highest prediction accuracy in the training set is defined as the Γ value.

2) Find the maximum value diff(K ₁ (x _m , _xi )) and minimum value diff(K ₁ (x _n , _xi )) of the tangent slope of the Gaussian kernel characteristic curve under the Γ value, and the corresponding tangent point coordinates (x _m , K ₁ (x _m , _xi )), (x _n , K ₁ (x _n , _xi )), this type of tangent point is called an inflection point.

3) According to the corner coordinates (x _m ，K ₁ (x _m , _xi )), (x _n ,K ₁ (x _n , _xi )) and slope diff(K ₁ (x _m , _xi ))/ ξ, diff(K ₁ (x _n , _xi ))/ξ, where ξ is a real number variable, determine the polynomial kernel expression, the expression is K ₂ '(x, _xi )=diff(K ₁ (x _m , x _i ))/ξ/ _xi ×(<x, _xi >+c) and K ₂ ″(x, _xi )＝diff(K ₁ (x _n , _xi ))/ξ/ _xi ×( <x, _xi >+c) consists of two parts.

4) Build a bionic kernel function

K _bsf (x,x _i )＝exp(-Γ||xx _i || ² )+diff(exp(-Γ||x _B -x _i || ² )/ξ/xi _× (<x,x _i >+c)

Where: when ||x||>||x _i ||, x _B is taken as x _n , when ||x||<||x _i ||, x _B is taken as x _m , ||x||＝ When ||x _i ||, classify x and _xi as the same class, and do not need to use the kernel function to calculate it.

Similar to the process in which the skin of a chameleon changes according to the color of the surrounding environment, the bionic kernel function K _bsf (x, _xi ) can adjust the global data and local data by adjusting the parameters Γ and ξ to adapt to the Requirements for different data classifications.

In fact, the bionic kernel function constructed above can be expressed as:

K _bsf (x, _xi )=K ₁ (x, _xi )+K ₂ (x, _xi ).

Among them: K ₁ (x, _xi ) is a Gaussian kernel, K ₂ (x, _xi ) is a polynomial kernel, according to the nature of the kernel function: if and only if the function K is negative definite, then K:X×X→R is the kernel function,

The proof process can be obtained as follows: for any a∈R ^l , we have Since K ₁ and K ₂ are kernel functions, so $a^T{K}_1^{*}a\&Greater\; Equal;0,a^T{K}_2^{*}a\&Greater\; Equal;0,$ can be pushed $a^T(K_1^{*}+K_2^{*})a\&Greater\; Equal;0,$ Right now * is positive semi-definite, so K ₁ +K ₂ is a kernel function, that is, the constructed K _bsf (x, _xi ) can be used as a kernel function.

4. Establish SVM training set and prediction set

Divide data M_D ₁ into two parts of data M_D ₁₁ and M_D ₁₂ according to different time periods, divide data M_D ₁ into two parts of data M_D ₁₁ and M_D ₁₂ according to the time period, and divide data C_D ₁ into two parts of data C_D according to corresponding time periods ₁₁ and C_D ₁₂ . The first part of data M_D ₁₁ and C_D ₁₁ is used as the training set of SVM, and the second part of data M_D ₁₂ and C_D ₁₂ is used as the prediction set of SVM.

5. Tool selection and parameter optimization for predictive model building

Referring to Fig. 3 and Fig. 4, utilize the SVM that has special advantage to small sample, put the bionic type kernel function that the present invention builds into libsvm toolbox, use M_D ₁₁ as SVM input quantity, C_D ₁₁ as SVM output quantity, for SVM The penalty parameter c, parameter γ in and the parameter Γ and parameter ξ in the bionic kernel function are optimized by grid-search. Using MATLAB, put the bionic kernel function K _bsf (x, _xi ) into the libsvm toolbox to replace the position of the original RBF kernel function, realize the grid-search method to optimize parameters through the function SVMcgForRegress(), and use the grid-search method to initially The parameter optimization process, as shown in Figure 3, adopts Gaussian kernelK ₁ (x, x _i ) in the function SVMcgForRegress (), through [bestmse, bestc, bestg]=SVMcgForRegress (C_D ₁₁ , M_D ₁₁ ,-8,8,- 8, 8), the parameters bestc and bestg can be obtained, bestc is the best penalty parameter c value, and bestg is the best parameter γ value, that is, the Γ value. As shown in Figure 4, the bionic kernel function K _bsf (x, x _i ) is used in the function SVMcgForRegress(), and [bestmse, bestc, bestCMG]=SVMcgForRegress(C_D ₁₁ , M_D ₁₁ , bestc, bestc,-8, 8), the best parameter bestCMG can be obtained, that is, the value of ξ.

6. Prediction of population net photosynthetic rate

1) Obtain the prediction model model

Using svmtrain() in the libsvm toolbox, that is, through model=svmtrain(C_D ₁₁ , M_D ₁₁ ,cmd), the prediction model model can be obtained, where cmd=['-c',num2str(bestc),'-g', num2str(bestg),'-s 3-p 0.01-t 2'], s is set to 3 to represent the e-SVR formula, p is set to 0.01 to set the value of the loss function in e-SVR, t is set to 2 to represent SVM The kernel function used in is K _bsf (x, _xi ); the model contains information about the type of kernel function, the number of support vectors and the coefficient range of support vectors in the decision function;

2) Obtain the predicted value predict of the net photosynthetic rate of the population

Using svmpredict() in the libsvm toolbox, through [predict,mse]=svmpredict(C_D ₁₂ ,M_D ₁₂ ,model) to realize the prediction of C_D _{12 using M_D 12 to} obtain the predicted value predict of the net photosynthetic rate of the population,

3) Judging the reliability of the prediction model model

At the same time, the correlation coefficient R between the two is calculated through C_D ₁₂ and predict, and the accuracy of the predicted value predict is judged, that is, the reliability of the prediction model model.

Example

1. Obtain the data of the spectral radiation ratio of each band of visible light in the region

Referring to Fig. 5, randomly select a test area of S= ^17m2 in the understory ginseng planting base, place the MSR-16 portable multi-spectral radiometer at the height H=2m from the understory ginseng plant canopy, and the observed diameter of the ground plants is H/2 = 1m, the observation area is S ₁ = 0.785, then S/S ₁ = 21.656, the number of test units M = 21, and the measurement period spans from July 1 to August 31, 2010. The measurement time For 9:00-16:00 on the same day of the test, there are 7 test time periods in total, so the number of test units selected is N=3, and the method for obtaining the spectral radiation ratio relationship data of each band of visible light in the region introduced by the present invention is used to obtain the data of each band of visible light in the region. The band spectral radiation ratio data M_D is normalized by [0, 1] to obtain the normalized data M_D ₁ , as shown in the figure.

2. Obtain population net photosynthetic rate data

Referring to Fig. 6, in each test unit in 21 test units, randomly select 3 strains of understory ginseng, and each plant randomly selects 5 blades, and adopts the population net photosynthetic rate data acquisition method introduced by the present invention to obtain understory ginseng. The net photosynthetic rate C_D of the population is normalized by [0, 1] to obtain the normalized data C_D ₁ , as shown in the figure.

3. Parameter optimization of the prediction model

The M_D ₁₁ from July 1st to August 10th in 2010 is taken as the SVM input, and the C_D ₁₁ during this period is taken as the SVM output. Utilize the optimization method to penalty parameter c and parameter γ that the present invention introduces, obtain optimum penalty parameter c=2.51, optimum parameter γ=1, namely Γ=1, MSE is 0.016; The function parameter optimization method finally determines the best parameter ξ of the bionic kernel function to be 1.68, and the parameter Γ to be 2.83.

4. Prediction of population net photosynthetic rate

Referring to Fig. 7 and Fig. 8, adopt the method for predicting the net photosynthetic rate of population that the present invention introduces, penalty parameter c=2.51, Γ=2.83, ξ=1.68, by model=svmtrain(C_D ₁₁ , M_D ₁₁ , cmd), obtain prediction model model, where model.totalSV is 146, that is, there are 146 support vectors, the minimum value in model.sv_coef is -0.625, and the maximum value is 0.625, that is, the coefficient range of the support vector in the decision function is [-0.625,0.625]. This model makes the best prediction accuracy of the training set C_D ₁₁ reach 89%, and the prediction effect is shown in Figure 7. Take M_D ₁₂ from August 11 to August 31, 2010 as the SVM input, and C_D ₁₂ during this period as the SVM output. Using this model, pass [predict,mse]=svmpredict(C_D ₁₂ ,M_D ₁₂ ,model), the prediction accuracy of the prediction set C_D ₁₂ reaches 81%, and the prediction effect is shown in Figure 8.

5. Judging the accuracy of the forecast results

Taking different S and H, repeating the above-mentioned group net photosynthetic rate prediction process, to obtain the group net photosynthetic rate prediction accuracy. The measurement time span is from July 1st to August 31st, 2010, and the measurement time is from 9:00 to 16:00 on the day of the test. The M_D ₁₁ and C_D ₁₁ is used as the training set of SVM, and M_D ₁₂ and C_D ₁₂ from August 11 to August 31, 2010 are used as the prediction set of SVM. When H=2m, (1) take S=6m ² , then S ₁ =0.785, S/S ₁ =7.643, the number of test units M=7, the number of test units N=1; (2) take S=11m ² , then S ₁ =0.785, S/S ₁ =14.006, the number of test units M=14, the number of test units N=2; (3) S=22m ² , then S ₁ =0.785, S/S ₁ =28.025, The number of test units M=28, the number of test units N=4; (4) S=28m ² , then S ₁ =0.785, S/S ₁ =35.669, the number of test units M=35, the number of test units N=5; (5) Take S=33m ² , then S ₁ =0.785, S/S ₁ =42.038, the number of test units M=42, the number of test units N=6; (6) Take S=39m ² , then S ₁ =0.785 , S/S ₁ =49.681, number of test units M=49, number of test units N=7; (7) S=44m ² , then S ₁ =0.785, S/S ₁ =56.051, number of test units M=56 , the number of test units N=8; (8) take S=50m ² , then S ₁ =0.785, S/S ₁ =63.694, the number of test units M=63, the number of test units N=9; (9) take S= 55m ² , then S ₁ =0.785, S/S ₁ =70.064, the number of test units M=70, and the number of test units N=10. The prediction accuracy of population net photosynthetic rate prediction set C_D ₁₂ is shown in Table 1.

Table 1H=2m, the prediction accuracy of the population net photosynthetic rate prediction set C_D ₁₂ under different S values

The measurement period spans from July 1 to August 31, 2011, and the measurement time is from 9:00 to 16:00 on the day of the test. The M_D ₁₁ and C_D from July 1 to August 10, 2011 ₁₁ is used as the training set of SVM, and M_D ₁₂ and C_D ₁₂ from August 11, 2011 to August 31, 2011 are used as the prediction set of SVM. When H=1.2m, (1) take S=2m ² , then S ₁ =0.283, S/S ₁ =7.067, the number of test units M=7, the number of test units N=1; (2) take S=4m ² , then S ₁ =0.283, S/S ₁ =14.134, the number of test units M=14, the number of test units N=2; (3) S=6m ² , then S ₁ =0.283, S/S ₁ =21.201 , the number of test units M=21, the number of test units N=3; (4) S=8m ² , then S ₁ =0.283, S/S ₁ =28.269, the number of test units M=28, the number of test units N=4 ; (5) S = 10m ² , then S ₁ = 0.283, S/S ₁ = 35.336, the number of test units M = 35, the number of test units N = 5; (6) S = 12m ² , then S ₁ = 0.283, S/S ₁ =42.403, number of test units M=42, number of test units N=6; (7) S=14m ² , then S ₁ =0.283, S/S ₁ =49.470, number of test units M= 49, the number of test units N=7; (8) take S=16m ² , then S ₁ =0.283, S/S ₁ =56.537, the number of test units M=56, the number of test units N=8; (9) take S =18m ² , then S ₁ =0.283, S/S ₁ =63.604, the number of test units M=63, the number of test units N=9; (10) S=20m ² , then S ₁ =0.283, S/S ₁ =70.671, the number of test units M=70, the number of test units N=10; the prediction accuracy of the net photosynthetic rate of the population is shown in Table 2. Table 2H＝1.2m, the prediction accuracy of population net photosynthetic rate under different S values

It can be seen from Table 1 and Table 2 that the net photosynthetic rate C_D of the population is predicted by using the bionic kernel function K _bsf (x,xi ₎ and the spectral radiation ratio data M_D of each band of visible light, and the prediction accuracy reaches 80%. As above, the effect is satisfactory, and the net photosynthetic rate of other plant populations in agriculture can be predicted by using a method for predicting the net photosynthetic rate of the population in the present invention.

Claims (3)

1. a group net photosynthetic rate prediction method, is characterized in that, the step of described a kind of group net photosynthetic rate prediction method is as follows: 1) Obtain the spectral radiation ratio data of each band of visible light in the region: (1) Install a portable multispectral radiometer model MSR-16 at a height H higher than the plant canopy in the test area with an area S, wherein: S takes a value of 6m ² to 55m ² , and H takes a value of 1.2m or 2m, the effective measurement diameter of the portable multispectral radiometer model MSR-16 on the ground directly below it is H/2, and the observation area is S ₁ , S ₁ =[π×(H/4) ² ], calculated The number of test units divided by the test area is M, and its value is the integer part of S/S ₁ ; (2) In the specified measurement time period, measure once every hour, randomly select N test units each time, where N<M, and require that the test units selected each time are different from the test units selected at other measurement times, each time When measuring, place the portable multi-spectral radiometer of MSR-16 at H above the test unit during the measurement. The average value of 5 fixed points is used as the spectral composition of the test unit, and the average value of 3 test units is used as the data of the spectral radiation ratio relationship of each band of visible light in the test area during this period; (3) According to the method of step (2), within the time span of the test, obtain the spectral radiation ratio data M_D of each band of visible light in the test area in different periods, and perform [0, 1] normalization processing to obtain normalization The transformed data M_D ₁ ; 2) Obtain the population net photosynthetic rate data: (1) The model of CI-310 portable photosynthesis measurement system produced by U.S. CID Company was adopted to measure the individual net photosynthetic rate; (2) Randomly select 3 plants of the same variety in each selected test unit. If the number of plants is less than 3, measure according to the actual number of plants. Each plant randomly selects 5 leaves. If the number of leaves is less than 5 , measured according to the actual number of leaves, and the net photosynthetic rate of each leaf is measured once with a portable photosynthesis measurement system model CI-310, and the average value is taken as the net photosynthetic rate of the plant, and the average value of the three test units is used as the test period. The net photosynthetic rate of the plant population in the area; (3) According to the method of step (2), within the test time span, obtain the net photosynthetic rate C_D of the plant population in the test area in different periods, carry out [0,1] normalization processing, and obtain the normalized data C_D ₁ ; 3) Build a bionic kernel function: (1) Based on Gaussian kernelK ₁ (x, _xi )=exp(-γ||xx _i || ² ) and polynomial kernelK ₂ (x, _xi )=k(<x, _xi >+c) Kernel function, where k, γ and c are parameters, x and _xi are multi-dimensional vectors in low-dimensional space, input K ₁ (x, _xi )=exp(-γ||xx _i || ² ) into libsvm, As the SVM kernel function, train the training set data, use grid-search to search for the best parameters, and determine the γ value with the highest prediction accuracy for the training set, which is defined as the Γ value; (2) Find out the maximum value diff(K ₁ (x _m , _xi )) and the minimum value diff(K ₁ (x _n , _xi )) of the tangent slope of the Gaussian kernel characteristic curve under the Γ value, and the corresponding tangent point Coordinates (x _m , K ₁ (x _m , _xi )), (x _n , K ₁ (x _n , _xi )), this type of tangent point is called an inflection point; (3) According to the corner coordinates (x _m ，K ₁ (x _m , _xi )), (x _n ,K ₁ (x _n , _xi )) and slope diff(K ₁ (x _m , _xi )) /ξ, diff(K ₁ (x _n , _xi ))/ξ, where ξ is a real number variable, determine the polynomial kernel expression, the expression is K ₂ '(x, _xi )=diff(K ₁ (x _m , _xi ))/ξ/ _xi ×(<x, _xi >+c) and K ₂ ″(x, _xi )＝diff(K ₁ (x _n ,xi ₎ )/ξ/xi _× (<x, _xi >+c) consists of two parts; (4) Build a bionic kernel function: K _bsf (x,x _i )＝exp(-Γ||xx _i || ² )+diff(exp(-Γ||x _B -x _i || ² )/ξ/xi _× (<x,x _i >+c) Where: when ||x||>||x _i ||, x _B is taken as x _n , when ||x||<||x _i ||, x _B is taken as x _m , ||x||＝ When ||x _i ||, classify x and _xi as the same class, and do not need to use the kernel function to calculate it; The bionic kernel function K _bsf (x, _xi ) can adjust the global data and local data through the adjustment of parameters Γ and ξ, so as to meet the requirements of different data classification; 4) Establish SVM training set and prediction set: Divide data M_D ₁ into two parts of data M_D ₁₁ and M_D ₁₂ according to different time periods, divide data M_D ₁ into two parts of data M_D ₁₁ and M_D ₁₂ according to the time period, and divide data C_D ₁ into two parts of data C_D according to corresponding time periods ₁₁ and C_D ₁₂ , the first part of data M_D ₁₁ and C_D ₁₁ are used as the training set of SVM, and the second part of data M_D ₁₂ and C_D ₁₂ are used as the prediction set of SVM; 5) Tool selection and parameter optimization for establishing predictive models; 6) Predict the net photosynthetic rate of the population. 2. according to a kind of group net photosynthetic rate prediction method according to claim 1, it is characterized in that, the tool selection and parameter optimization of described establishment prediction model refer to: Using MATLAB to put the bionic kernel function K _bsf (x, _xi ) into the libsvm toolbox to replace the position of the original RBF kernel function, realize the grid-search method to optimize parameters through the function SVMcgForRegress(), and use the grid-search method for preliminary optimization Parameter process, using Gaussian kernelK ₁ (x, _xi ) in the function SVMcgForRegress(), through [bestmse,bestc,bestg]=SVMcgForRegress(C_D ₁₁ ,M_D ₁₁ ,-8,8,-8,8), you can get Parameters bestc and bestg, bestc is the best penalty parameter c value, bestg is the best parameter γ value, that is, Γ value; _The bionic kernel function K _bsf (x, _{xi )} is used in the function SVMcgForRegress(), and the best The best parameter bestCMG is the value of ξ. 3. according to a kind of group net photosynthetic rate predicting method according to claim 1, it is characterized in that, the step of described prediction group net photosynthetic rate is as follows: 1) Obtain the prediction model model Using svmtrain() in the libsvm toolbox, that is, through model=svmtrain(C_D ₁₁ , M_D ₁₁ ,cmd), the prediction model model can be obtained, where cmd=['-c',num2str(bestc),'-g', num2str(bestg),'-s 3-p 0.01-t 2'], s is set to 3 to represent the e-SVR formula, p is set to 0.01 to set the value of the loss function in e-SVR, t is set to 2 to represent SVM The kernel function used in is K _bsf (x, _xi ); the model contains information about the type of kernel function, the number of support vectors and the coefficient range of support vectors in the decision function; 2) Obtain the predicted value predict of the net photosynthetic rate of the population Utilize svmpredict() in the libsvm toolbox, through [predict,mse]=svmpredict(C_D ₁₂ ,M_D ₁₂ ,model) to realize using M_D ₁₂ to predict C_D ₁₂ , and obtain the predicted value predict of the net photosynthetic rate of the population; 3) Judging the reliability of the prediction model model At the same time, the correlation coefficient R between the two is calculated through C_D ₁₂ and predict, and the accuracy of the predicted value predict is judged, that is, the reliability of the prediction model model.